Neurosymbolic Reinforcement Learning with Formally Verified Exploration